The Hague, Netherlands August 30, 2016
Proceedings of the 2 nd Workshop on
Artificial Intelligence and Internet of Things (2 nd AI-IoT 2016)
In conjunction with
ECAI 2016
The Hague, Netherlands August 30, 2016
Proceedings of the 2 nd Workshop on
Artificial Intelligence and Internet of Things (2 nd AI-IoT 2016)
In conjunction with
ECAI 2016
Organization
Chair
Constantine D. Spyropoulos NCSR “Demokritos”, Greece
Program Committee
Constantine D. Spyropoulos NCSR “Demokritos”, Greece
Alexander Artikis NCSR “Demokritos” & University of Piraeus, Greece Payam Barnaghi University of Surrey, United Kingdom
Amedeo Cesta National Research Council of Italy, Italy Alfio Ferrara University of Milano, Italy
Amelie Gyrard Insight Centre for Data Analytics, Ireland Alexander Jungmann University of Paderborn, Germany Konstantinos Kotis University of Piraeus, Greece
Katerina Marinagi Technological Educational Institute of Chalkis, Greece Stavros Perantonis NCSR “Demokritos”, Greece
Georgios Pierris NCSR “Demokritos”, Greece Evangelos Pournaras ETH Zurich, Switzerland Ioannis Refanidis University of Macedonia, Greece Yannis Soldatos Athens Information Technology, Greece Evaggelos Spyrou NCSR “Demokritos”, Greece
Martin Strohbach AGT International, Switzerland
Grigorios Tsoumakas Aristotle University of Thessaloniki, Greece Grigorios Tzortzis NCSR “Demokritos”, Greece
Nick Vassiliadis Aristotle University of Thessaloniki, Greece Ioannis Vlahavas Aristotle University of Thessaloniki, Greece George Vouros University of Piraeus, Greece
Organizing Committee
Constantine D. Spyropoulos NCSR “Demokritos”, Greece Georgios Pierris NCSR “Demokritos”, Greece Grigorios Tzortzis NCSR “Demokritos”, Greece
Table of Contents
Preface . . . v
Keynote . . . vii
Third Generation Teleassistance: Intelligent Monitoring Makes the
Difference. . . 1 Xavier Rafael-Palou, Carme Zambrana, Stefan Dauwalder, Enrique
de la Vega, Eloisa Vargiu and Felip Miralles
Temporal Goal Reasoning for Predictive Performance of a Tourist
Application . . . 7 Eliseo Marzal, Jesus Iba˜nez, Laura Sebastia and Eva Onaindia
An Intelligent System for Smart Tourism Simulation in a Dynamic
Environment . . . 15 Mohannad Babli, Jesus Iba˜nez, Laura Sebastia, Antonio Garrido
and Eva Onaindia
Extending Naive Bayes with Precision-tunable Feature Variables for
Resource-efficient Sensor Fusion. . . 23 Laura Isabel Galindez Olascoaga, Wannes Meert, Herman
Bruyninckx and Marian Verhelst
A Distributed Event Calculus for Event Recognition. . . 31 Alexandros Mavrommatis, Alexander Artikis, Anastasios Skarlatidis
and Georgios Paliouras
Preface
The continuous international efforts to enable everyday devices to partici- pate in the emerging Internet of Things (IoT) ecosystem has led to an explosive increase in the number of smart devices that surround us. However, our capacity as humans to meaningfully process, manage, control, and interact with them is limited by human nature, our interests and our technical fluency. The coming new digital market envisions an ambient environment where the physical world, computer-based systems and humans converge and seamlessly interoperate, re- sulting in an improved social and economic marketplace.
Collectively, the public sector, industry, academia, end-users, SMEs, and large corporations constantly feed the, already, high expectations of IoT. Artifi- cial Intelligence (AI) has the capacity to facilitate the anticipated socio-economic transformation caused by the proliferation of IoT through innovative algorithms and techniques.
The Artificial Intelligence and Internet of Things (AI-IoT) series of work- shops aims at providing the ground for disseminating new and interesting ideas on how AI can make valuable contribution in solving problems that the IoT ecosystem faces. The virtualization of devices and smart systems, the discover- ability and composition of services, the interoperability of services, the distribu- tion of resources, the management and event recognition of big stream data, and the development of algorithms for edge and predictive analytics are only a few of the problems that look for intelligent human-centric solutions that could find application in smart cities, smart farming, transportation, health, smart grid, tourism, etc.
The second installment of the workshop – 2nd AI-IoT 2016 – was co-located with ECAI 2016 in The Hague, Netherlands and featured a keynote by Prof. Dirk Helbing from ETH Zurich, Switzerland, entitled “Towards Smarter Societies”
and five accepted papers, resulting in an intriguing technical program. Papers accepted in the workshop gave special emphasis in AI-related topics such as:
– Machine learning – AI planning
– Reasoning under uncertainty – Personalization
– Classification
– Real-time event recognition – Multi-agent systems
that have been explored in smart societies, tele-assistance, smart tourism, em- bedded sensor fusion, for activity recognition in surveillance and security sys- tems, and for detecting treads and abnormal activities in maritime surveillance.
Specifically in this proceedings the contributions of the accepted papers are as follows. In the “Third Generation Teleassistance: Intelligent Monitoring Makes
the Difference”, Rafael-Palou et al. propose an intelligent monitoring solution for elderly people, integrated in an IoT-based tele-assistance system, demon- strating how it contributes in offering enhanced support to both end-users and caregivers. Machine learning methods based on SVM are used for detecting in- teresting events and issuing alarms in case of an emergency. Results from deploy- ing the system in real-life situations are presented. Marzal et al. in “Temporal Goal Reasoning for Predictive Performance of a Tourist Application”, discuss a goal reasoning framework that identifies if the context information acquired from several external resources dictates a change in the execution of a temporal plan. TempLM, a temporal planner that uses temporal landmarks for planning with temporal deadlines, detects situations of future failures and opportunities in the plan execution. The capability of the planner to adapt to external events is showcased in a smart tourism scenario. Babli et al. in their paper entitled
“An Intelligent System for Smart Tourism Simulation in a Dynamic Environ- ment” present an AI planning-based system for the smart tourism domain, where the goal is to construct a personalized tourist agenda of places a tourist could visit according to his preferences. The system not only creates the agenda, but also monitors its execution in real-time through simulation. Emphasis is given in dynamically reacting to changes in the environment by adapting, if neces- sary, the tourist agenda, through reformulation of the planning problem, to re- flect the new state of the environment in real-time. In “Extending Naive Bayes with Precision-tunable Feature Variables for Resource-efficient Sensor Fusion”, Galindez Olascoaga et al. focus on the tradeoff between resource efficiency and inference accuracy, by tuning feature quality in sensing devices. An extension to the naive Bayes classifier is implemented and evaluated in sensor fusion tasks.
The algorithm is capable of dynamically tuning feature precision as a function of the incoming data quality, the difficulty of the task and the resource avail- ability. In the last paper, “A Distributed Event Calculus for Event Recognition”, Mavrommatis et al. present a distributed approach for stream reasoning, called dRTEC, based on a dialect of event calculus. dRTEC employs the Apache Spark framework to perform scalable event recognition and detect significant patterns.
The organizers would like to thank the authors for submitting their work to the workshop, the members of the program committee for their valuable contribution in reviewing the papers and, of course, the numerous participants of the workshop.
Constantine D. Spyropoulos Georgios Pierris Grigorios Tzortzis Organizers of 2nd AI-IoT 2016 Workshop site: http://2nd-ai-iot2016.iit.demokritos.gr/
Keynote
Speaker:Prof. Dirk Helbing, ETH Zurich, Switzerland Title:Towards Smarter Societies
Abstract – As the recent triumph of alphaGo has shown, the exponential in- crease in computer power allows us now to solve challenging problems that seemed to be out of reach for a long time. So, would we eventually be able to build superintelligent computers that would be able to solve humanities’ 21st century problems and run our society in an optimal way? Surprisingly, the an- swer is “no”, because data volumes increase faster than processing power, and systemic complexity even faster. This has a number of implications: local knowl- edge, context as well as distributed computing and control will become more important. Science will be relevant again to decide what data to process and how, and how to collect the right kind of data in the first place. This talk will elaborate on a number of pitfalls in the areas of Data Science and AI, and it will make proposals how to use these technologies and the Internet of Things more successfully, with context-aware approaches.
Third Generation Teleassistance:
Intelligent Monitoring Makes the Difference
Xavier Rafael-Palou
1and Carme Zambrana
1and Stefan Dauwalder
1and Enrique de la Vega
2and Eloisa Vargiu
1and Felip Miralles
1Abstract. Elderly people aim to preserve their independence and autonomy at their own home as long as possible. However, as they get old the risks of disease and injuries increase making critical to assist and provide them the right care whenever needed. Unfortu- nately, neither relatives, private institutions nor public care services are viable long-term solutions due to the large amount of required time and cost. Thus, smart teleassistance solutions must be investi- gated. In particular, IoT paradigm helps in designing third generation teleassistance systems by relying on sensors to gather the more data as possible. Moreover, we claim that providing IoT solutions of in- telligent monitoring improves the overall efficacy. In this paper, we presents an intelligent monitoring solution, fully integrated in a IoT- based teleassistance system, showing how it helps in giving better support to both end-users and carers. Thanks to intelligent monitor- ing, carers can instantly access to the relevant information regard- ing the status of the end-user, also receiving alarms in case of any anomaly or emergency situations have been detected.
1 Introduction
In the last decade, the Internet of Things (IoT) paradigm rapidly grew up gaining ground in the scenario of modern wireless telecommu- nications [6]. Its basic idea is the pervasive presence of a variety of things or objects (e.g, tags, sensors, actuators, smartphones, everyday objects) that are able to interact with each other and cooperate with their neighbors to reach common goals. IoT solutions have been in- vestigated and proposed in several fields [7], such as automotive [17], logistics [19], agriculture [31], entertainment [18], and independent living [12].
Several research issues are still open: standardization, networking, security, and privacy [27]. We claim that research might also focus on intelligent techniques to improve IoT solutions thus making the difference with respect to classical systems. In other words, artificial intelligence algorithms and methods may be integrated in IoT sys- tems: to allow better coordination and communication among sen- sors, through adopting multi-agent systems [1]; to adapt the sensor network according to the context, by relying, for instance, on deep learning techniques [16]; as well as to provide recommendations to the final users, by using data fusion and semantic interpretation [4].
Considering the dependency care sector as a case study, in this paper we show how intelligent monitoring techniques, integrated in
1 eHealth Unit, EURECAT, Barcelona, email: {xavier.rafael, carme.zambrana, stefan.dauwalder, eloisa.vargiu, fe- lip.miralles}@eurecat.org
2 Technology Transfer Unit, EURECAT, Barcelona, email: en- rique.delavega@eurecat.org
a IoT-based teleassistance system (namely, eKauri3), help in pro- viding better assistance and support to people that need assistance.
eKauri is a teleassistance system composed of a set of wireless sen- sors connected to a gateway (based on Raspberry-pi) that collects and securely redirects them to the cloud. It is worth noting that eKuari is composed by the following kinds of sensors: one presence- illumination-temperature sensors (i.e., TSP01 Z-Wave PIR) for each room, and one presence-door-illumination-temperature sensor (i.e., TSM02 Z-Wave PIR) for each entry door. Intelligent monitoring in eKauri allows to detect the following events: leaving home; going back to home; receiving a visit; remaining alone after a visit; go- ing to the bathroom; going to sleep; and awaking from sleep. In this paper, we focus on the contribution of the intelligent monitoring in eKauri, the interested reader may refer to [23] for a deep description of the system.
The rest of the paper is organized as follows. In Section 2, we briefly recall IoT solutions to teleassistance. Section 3 illustrates how intelligent monitoring improves teleassistance in the eKauri system.
In Section 4, the main installations of eKauri are presented together with users’ experience. Section 5 ends the paper summarizing the main conclusions.
2 Related Work
Teleassistance remotely, automatically and passively monitors changes in people’s condition or lifestyle, with the final goal of man- aging the risks of independent living [9] [2]. In other words, thanks to teleassistance, end-users are connected with therapists and care- givers as well as relatives and family, allowing people with special needs to be independent.
There are several of efforts to utilize IoT-based systems for mon- itoring elderly people, most of which target only certain aspects of elderly requirements from a limited viewpoint. Gokalp and Clarke reviewed monitoring activities of daily living of elderly people com- paring characteristics, outcomes, and limitations of 25 studies [15].
They found that adopted sensors are mainly environmental, except for accelerometers and some physiological sensors. Ambient sensors could not differentiate the subject from visitors, as opposed to wear- able sensors [8] [5]. On the other hand, the latter could only distin- guish simple activities, such as walking, running, resting, falling, or inactivity [3]. Moreover, wearable sensors are not suitable for cogni- tively impaired elderly people due to the fact that they are likely to be forgotten or thrown away [11] [14]. Their main conclusion regard- ing sensors is that daily living activity monitoring requires use of a combination of ambient sensors, such as motion and door sensors.
3www.ekauri.com
3 Intelligent Monitoring Makes the Difference
Filtering and analyzing data coming from teleassistance systems is becoming more and more relevant. In fact, a lot of data are continu- ously gathered and sent through the sensors. The role of therapists, caregivers, social workers, as well as relatives (hereinafter, carers) is essential for remotely assisting monitored users. On the one hand, the monitored user (e.g., elderly or disabled people) needs to be kept informed about emergencies as soon as they happen and s/he has to be in contact with therapists and caregivers to change habits and/or to perform some therapy. On the other hand, monitoring systems are very important from the perspective of carers. In fact, those systems allow them to become aware of user context by acquiring heteroge- neous data coming from sensors and other sources. Thus, intelligent solutions able to understand all those data and process them to keep carers aware about their assisted persons are needed, providing also users empowerment.
In the following, we show how intelligent monitoring helps in: im- proving sensors reliability allowing better activity recognition; pro- viding useful information to carers; and inferring quality of life of users.
3.1 Improving Sensors Reliability
Performance of IoT systems depends, among other characteristics, on the reliability of the adopted sensors. In the case of teleassistance, binary sensors are quite used in the literature and also in commercial solutions to identify user’s activities. Binary sensors do not have the ability to directly identify people and can only present two possible values as outputs (“0” and “1”). Typical examples of binary sensors deployed within smart environments include pressure mats, door sen- sors, and movement detectors. A number of studies reporting the use of binary and related sensors have been undertaken for the purposes of activity recognition [26]. Nevertheless, sensor data can be consid- ered to be highly dynamic and prone to noise and errors [25]. In the following, we present two solutions that rely on machine learning to improve reliability of sensors in presence detection and sleeping recognition, respectively.
3.1.1 Presence Detection
Detecting user’s entering/leaving home can be done by relying on door sensors. Fusing data from door- and motions-sensors could help also in recognizing if the user received visits. Unfortunately, as said, sensors are not 100% reliable: sometimes they loose events or detect them several times. When sensors remain with a low battery charge they get worse. Moreover, also the Raspberry pi may loose some data or the connection with Internet and/or with the sensors. Also the In- ternet connection may stop working or loose data. Finally, without using a camera or wearable sensors we are not able to directly recog- nize if the user is alone or if s/he has some visits.
In order to solve this kind of limitations with the final goal of im- proving the overall performance of our IoT-based system that uses only motion and door sensors, we defined and adopt a two-levels hi- erarchical classifier (see Figure 1) [24]: the upper level is aimed at recognizing if the user is at home or not, whereas the lower is aimed at recognizing if the user is really alone or if s/he received some vis- its.
The goal of the classifier at the upper level is to improve perfor- mance of the door sensor. In fact, it may happen that the sensor reg- isters a status change (from closed to open) even if the door has not
Figure 1. The hierarchical approach to presence detection.
been opened. This implies that the system may register that the user is away and, in the meanwhile, activities are detected at user’s home.
On the contrary, the system may register that the user is at home and, in the meanwhile, activities are not detected at user’s home. To solve, or at least reduce, this problem, we built a supervised classifier able to recognize if the door sensor is working well or erroneous events have been detected. First, we revise the data gathered by the sensor- based system searching for anomalies, i.e.: (1) the user is away and at home some events are detected and (2) the user is at home and no events are detected. Then, we validated those data by relying on Moves, an app installed and running on the user smartphone4. In fact, Moves, among other functionality, is able to localize the user. Hence, using Moves as an “oracle” we build a dataset in which each entry is labeled depending on the fact that the door sensor was right (label
“1”) or wrong (label “0”).
The goal of the classifier at the lower level is to identify whether the user is alone or not. The input data of this classifier are those that has been filtered by the upper level, being recognized as positives. To build this classifier, we rely on the novelty detection approach [20]
used when data has few positive cases (i.e., anomalies) compared with the negatives (i.e., regular cases); in case of skewed data.
The hierarchical approach was part of the EU project BackHome5. To train and test it, we consider a window of 4 months for training and evaluation (training dataset) and a window of 1 month for the test (testing dataset). Experiments have been performed at each level of the hierarchy. First, we performed experiments to identify the best supervised classifier to be used at the upper level of the hierarchy.
The best performance has been obtained by relying on the SVM (with γ = 1.0and C = 0.452). Subsequently, we applied the novelty detection algorithm on the data filtered by the classifier at the upper level, to validate the classifier at the lower one. Finally, we measure the performance of the overall approach. We compared the overall results with those obtained by using the rule-based approach in both levels of the hierarchy. Results are shown in Table 1 and point out
4https://www.moves-app.com/
5www.backhome-fp7.eu
that the proposed approach outperforms the rule-based one with a significant improvement.
Table 1. Results of the overall hierarchical approach with respect to the rule-based one.
Metric Rule-based Hierarchical Improv.
Accuracy 0.80 0.95 15%
Precision 0.68 0.94 26%
Recall 0.71 0.91 20%
F1 0.69 0.92 23%
3.1.2 Sleep Recognition
We defined the sleeping activity as the period which begins when the user goes to sleep and ends when the user wakes up in the morning.
Sleep recognition is aimed at reporting the following information: (i) the time when the user went to sleep and woke up; hereinafter we will refer to them asgo to sleep timeandwake up time, respectively;
(ii) the number of sleeping activity hours; and (iii) the number of rest hours, which are sleeping activity hours minus the time that the user spent going to the toilet or performing other activities during the night.
Let us note that the simplest way to recognize sleeping activities is relying on a rule-based approach. In particular, the following rules may be adopted: the user is in the bedroom; the activity is performed at night (e.g., the period between 8 pm to 8 am); the user is inactive;
and the inactivity duration is more than half an hour. Unfortunately, when moving to the real-world, some issues arise: user movements in the bed might be wrongly classified as awake; rules assumed all users wake up before 8 A.M., which is a strong assumption; and the ap- proach cannot distinguish if the user is, for instance, in the bedroom watching TV or reading a book, thus classifying all those actions as sleeping.
In order to overcome those limitations, an SVM (Radial Basis Function kernel, withC = 1.0,γ = 1.0) has been adopted to clas- sify the periods between two bedroom motions in two classes, awake and sleep [32]. Let us note that awake corresponds to the period in which the user goes to another room; performs activities in the bed- room; or stays in the bedroom with the light switched on. Otherwise, the activity is sleep.
Experiments, performed from May 2015 to January 2016 in 13 homes in Barcelona, show that the adopted machine learning solu- tion is able to recognize when the user is performing her/his sleeping activity. In particular, the proposed approach reaches an F1 of 96%.
Moreover, the adopted classifier is able to easily detect thego to sleep time, thewake up time, the number of sleeping activity hours and the number of rest hours. Figure 2 shows the comparisons between the ground truth (obtained by questionnaires answered by the users) and the results obtained with the machine learning approach (based on an SVM classifier). The plot has as temporal axis (axis x) and each co- ordinate in axis y represents nights in the dataset. The figure shows, in red, the sleep activity hours according to the ground truth and, in blue, the sleep activity hours calculated by the system. As both sleep activity hours of the same night are plotted in the same y coordinate, if the ground truth and the results coincide the color turns purple. If thego to sleep timeand/orwake up timedo not coincide, there is a text next to the corresponding side with the difference between the time coming from the ground truth and that coming from the results.
In the middle of each bar there is the total time which results differ from the baseline.
Figure 2. Comparison between the ground truth and the machine-learning (SVM) one.
3.2 Providing Feedback to Carers
The role of carers is essential for remotely assisting people that need assistance. Thus, intelligent monitoring able to understand gathered data and process them to keep carers aware about their assisted per- sons are needed [13].
Figure 3. The main information given to carers through the healthcare center.
Thanks to the user-centered approach from the above-mentioned projects, we designed friendly and useful interfaces for accessing and visualizing relevant data and information. In particular, carers iden- tified as the most relevant the following information (see Figure 3, first line on the top): time spent making activities, time spent sleep- ing, number of times the user leaves the home (during both day and night), and number of times the user goes to toilet (during both day and night). Moreover, they considered relevant to visually show the rooms where the user stayed time after time during a day (see Figure 3, central part) or during a period (e.g., the last month, as shown in Figure 4). They also want to be informed about all the notifications, chronologically ordered (see Figure 3, on the bottom). Finally, they want to access to some statistics to be aware about the evolution of user’s habits in order to act accordingly.
Figure 4. 1 month reporting.
To highlight the relevance of providing suitable information to car-
ers, let us mention here two cases that happened during Barcelona in- stallations in collaboration with Centre de Vida Independent6.Case- 1.A woman with Alzheimer and heart problems needs continuously assistance and, thus, a caregiver visits her daily. One day, eKauri de- tected that no visits were received, an alarm was generated and the caregiver called. The caregiver confirmed that she did not go to visit the user that day.Case-2.During the afternoon, a user is accustomed to go out for a walk. One day, she stayed in the bedroom. eKauri de- tected the change in her habit and a caregiver called her. Actually, she had a problem with a knee and she could not walk. A physiotherapist was asked to go to visit her.
3.3 Assessing Quality of Life of Users
In the dependency care sector, analyzing data gathered by sensors may help in improving teleassistance systems in becoming aware of user context. In so doing, they would be able to automatically in- fer user’s behavior as well as detect anomalies. In this direction, we studied a solution aimed at automatically assessing quality of life of people [29]. The goal is twofold: to provide support to people in need of assistance and to inform therapists, carers and families about the improvement/worsening of quality of life of monitored people.
First, we defined a Visual Analogic Scale (VAS) QoL question- naire composed of the following items:MOOD,HEALTH,MOBIL- ITY,SATISFACTION WITH CARE,USUAL ACTIVITIES(which in- cludes SLEEPING), and PAIN/DISCOMFORT. Those items have been categorized in two families: monitorable and inferable. Mon- itorable items can be directly gathered from sensors without relying on direct input from the user. Inferable items can be assessed by an- alyzing data retrieved by the system when considering activities per- formed by the user not directly linked with the sensors.
We performed experiments on two monitorable items (i.e.,MO- BILITYandSLEEPING) and one inferable (i.e.,MOOD). In partic- ular, we are able to detect and acknowledge the location of the user over time as well as the covered distance in kilometers and the places where s/he stayed. At the same time, we can detect when the user is sleeping as well as how many times s/he is waking up during the night. Merging and fusing the information related toMOBILITYand SLEEPING, we may also infer the overallMOOD.
The corresponding QoL assessment system is composed of a set of sub-modules, each one devoted to assess a specific QoL item;
namely: MOBILITY-assessment module; SLEEPING-assessment module; andMOOD-assessment module. Each sub-module is com- posed of two parts: Feature Extractor and Classifier. The Feature Ex- tractor receives as input the list of notifications{n}and the list of ac- tivities{a}and extracts the relevant features{f}to be given as input to the Classifier. The Classifier, then, uses those features to identify the right class Cl. This information will be then part of the overall summaryΣ.
Each Feature Extractor works with its proper list of features:
• MOBILITY: number of times the user left home, total time per- forming outdoor activities, total time performing activities (both indoors and outdoors), total time of inactivity, covered distance, number of performed steps, number of visited places, number of burned calories.
• SLEEPING: total sleeping time, hour the user went to sleep, hour the user woke up, number of times the user went to the toilet dur- ing the night, time spent at the toilet during the night, number of time the user went to the bedroom during the night, time spent at
6http://www.cvi-bcn.org/en/
the bedroom during the night, number of sleeping hours the day before, number of sleeping hours in the five days before.
• MOOD: number of received visits, total time performing outdoor activities, total time performing activities (both indoors and out- doors), total time of inactivity, covered distance, number of per- formed steps, number of burned calories, hour the user went to sleep, hour the user woke up, number of times the user went to the toilet during the night, time spent at the toilet during the night, number of time the user went to the bedroom during the night, time spent at the bedroom during the night, number of sleeping hours the day before, number of sleeping hours in the five days before. The Classifier is a supervised multi-class classifier built by using data previously labeled by the user and works on five classes, Very Bad, Bad, Normal, Good, and Very Good.
Under the umbrella of BackHome, we tested our approach with 3 users with severe disabilities (both cognitive and motor) living at their own real homes [30]. Although the system was evaluated by using as ground truth answers given to QoL questionnaires that is an approach completely subjective that depends on the particularity of each monitored user, after only 3 weeks of testing, the approach seemed convincing. Results presented in this paper show thatMO- BILITY,SLEEPING, andMOODcan be inferred with a high accu- racy (0.76, 0.72, and 0.81, respectively) by relying on an automatic QoL assessment system. Let us note thatSLEEPINGwas the method with the lowest performance. This is due to the fact that, currently, the system uses only motion sensors. Higher performances could be expected when combining motion sensors with other ones, such as mat-pressure or light sensors.MOBILITY achieved higher perfor- mance results thanSLEEPINGespecially when outdoor and indoor features are merged together. In fact, using only outdoor features was not as reliable as combining with indoor. This can be due to the re- liability of the GPS system embedded in the smartphone that made some errors in identifying when the user was really away. Let us also note that this is an important result because disable people in gen- eral spend a lot of time at their home. Finally,MOODreported the highest performances. Although at a first instance this could be sur- prising, this fact might be explained considering the intrinsic correla- tion betweenSLEEPINGandMOBILITY, as highlighted by the ques- tionnaire compiled daily by the users. It is worth noting that higher performances could be expected considering also social networking activities performed by the user.
4 Users’ Experience
The proposed solution has been developed according to a user- centered design approach in order to collect requirements and feed- back from all the actors (i.e., end-users and their relatives, profes- sionals, caregivers, and social workers). For evaluation purposes, the system has been installed in two healthy-user homes in Barcelona (control users).
The system has been used in the EU project BackHome to monitor disabled people. BackHome was an European R&D project that fo- cuses on restoring independence to people that are affected by motor impairment due to acquired brain injury or disease, with the over- all aim of preventing exclusion [21] [22]. In BackHome, informa- tion gathered by the sensor-based system is used to provide context- awareness by relying on ambient intelligence [10]. Intelligent mon- itoring was used in BackHome to study habits and to automatically assess QoL of people. The BackHome system ran in 3 end-user’s home in Belfast.
In collaboration with Centre de Vida Independent7, from May 2015 to January 2016, eKauri was installed in Barcelona in 13 el- derly people’ homes (12 women) over 65 years old [28]. To test eKauri, monitored users were asked to daily answer to a question- naire composed of 20 questions (12 optional). Moreover, they daily received a phone-call by a caregiver who manually verifies the data.
All detected events were shown in the Web applications and revised by therapists and caregivers. Feedback from them has been used to improve the interface and add functionality.
Although, at least at the beginning, users were a little bit reticent, during the monitored period they felt comfortable with the services provided by eKauri. In particular, they really appreciated the fact that it is not-intrusive and that it allows them to follow their normal lives.
In the case of CVI, people also be grateful for being called by phone.
In other words, it is important to provide a system that may become part of the home without losing social interactions. Thus, a teleassis- tance system does not substitute the role of caregivers. On the other side, carers recognized eKauri as a support to detect users’ habits helping in diagnosing user’s conditions and her/his decline, if any.
Currently, eKauri is installed in 40 elderly people’s homes in the Basque Country in collaboration with Fundaci´on Salud y Comu- nidad8.
5 Conclusions
Considering the dependency care sector as a case study, in this paper we highlighted how intelligent monitoring techniques, integrated in eKauri, an IoT-based teleassistance system, allow to better provide assistance and support to people that need assistance. In particular, we focused on the power of intelligent monitoring in improving sen- sor reliability, activity recognition, feedback provided to carers, as well as quality of life of final users. As a matter of fact, results about independent home evaluation of eKauri show a good acceptance of the system by both home users and caregivers. Being promising, the potential socio-economic impact of the exploitation of the system, as well as barriers and facilitators for future deployment, have to be analyzed before going to the market.
Summarizing, our main conclusion is that time is ripe to adopt IoT in the real world and that intelligent monitoring makes the difference in providing feedback to the users. However, to become pervasive, in particular in the dependency care sector, solutions must be taken into account the role of the final users in each phase of the development.
Moreover, even if users at home and caregivers give a positive feed- back, one step ahead might be performed to allow that stakeholders will take value from third generation teleassistance systems. It means that, as technological providers, we must put into effect concrete so- lutions that give a twist in adapting innovative strategies.
Acknowledgments
This research was partially funded by the European Community, un- der the BackHome project (grant agreement n. 288566 - Seventh Framework Programme FP7/2007-2013), and the CONNECARE project (grant agreement n. 689802 - H2020-EU.3.1). This paper re- flects only the authors’ views and funding agencies are not liable for any use that may be made of the information contained herein.
7http://www.cvi-bcn.org/en/
8https://www.fsyc.org/
REFERENCES
[1] Charilaos Akasiadis, Evaggelos Spyrou, Georgios Pierris, Dimitris Sgouropoulos, Giorgos Siantikos, Alexandros Mavrommatis, Costas Vrakopoulos, and Theodoros Giannakopoulos, ‘Exploiting future in- ternet technologies: the smart room case’, inProceedings of the 8th ACM International Conference on PErvasive Technologies Related to Assistive Environments, p. 85. ACM, (2015).
[2] Stelios Andreadis, Thanos G Stavropoulos, Georgios Meditskos, and Ioannis Kompatsiaris, ‘Dem@ home: Ambient intelligence for clinical support of people living with dementia’, inProc. 1st Workshop on Se- mantic Web Technologies in Pervasive and Mobile Environments (SEM- PER2016), (2016).
[3] Urs Anliker, Jamie A Ward, Paul Lukowicz, Gerhard Tr¨oster, Francois Dolveck, Michel Baer, Fatou Keita, Eran B Schenker, Fabrizio Catarsi, Luca Coluccini, et al., ‘Amon: a wearable multiparameter medical mon- itoring and alert system’,Information Technology in Biomedicine, IEEE Transactions on,8(4), 415–427, (2004).
[4] Alexander Artikis, Panagiotis D Bamidis, Antonis Billis, Charalampos Bratsas, Christos Frantzidis, Vangelis Karkaletsis, Manousos Klados, Evdokimos Konstantinidis, Stasinos Konstantopoulos, Dimitris Kos- mopoulos, et al., ‘Supporting tele-health and ai-based clinical decision making with sensor data fusion and semantic interpretation: The use- fil case study’, inInternational Workshop on Artificial Intelligence and NetMedicine, p. 21, (2012).
[5] Louis Atallah, Benny Lo, Raza Ali, Rachel King, and Guang-Zhong Yang, ‘Real-time activity classification using ambient and wearable sensors’,Information Technology in Biomedicine, IEEE Transactions on,13(6), 1031–1039, (2009).
[6] Luigi Atzori, Antonio Iera, and Giacomo Morabito, ‘The internet of things: A survey’,Computer networks,54(15), 2787–2805, (2010).
[7] Debasis Bandyopadhyay and Jaydip Sen, ‘Internet of things: Applica- tions and challenges in technology and standardization’,Wireless Per- sonal Communications,58(1), 49–69, (2011).
[8] NP Bidargaddi, A Sarela, et al., ‘Activity and heart rate-based mea- sures for outpatient cardiac rehabilitation’,Methods of information in medicine,47(3), 208–216, (2008).
[9] Peter Bower, Martin Cartwright, Shashivadan P Hirani, James Barlow, Jane Hendy, Martin Knapp, Catherine Henderson, Anne Rogers, Car- oline Sanders, Martin Bardsley, et al., ‘A comprehensive evaluation of the impact of telemonitoring in patients with long-term conditions and social care needs: protocol for the whole systems demonstrator cluster randomised trial’,BMC health services research,11(1), 184, (2011).
[10] Eloi Casals, Jos´e Alejandro Cordero, Stefan Dauwalder, Juan Manuel Fern´andez, Marc Sol`a, Eloisa Vargiu, and Felip Miralles, ‘Ambient in- telligence by atml: Rules in backhome’, inEmerging ideas on Informa- tion Filtering and Retrieval. DART 2013: Revised and Invited Papers;
C. Lai, A. Giuliani and G. Semeraro (eds.), (2014).
[11] Marie Chan, Eric Campo, and Daniel Est`eve, ‘Assessment of activity of elderly people using a home monitoring system’,International Journal of Rehabilitation Research,28(1), 69–76, (2005).
[12] Angelika Dohr, R Modre-Opsrian, Mario Drobics, Dieter Hayn, and G¨unter Schreier, ‘The internet of things for ambient assisted living’, in 2010 Seventh International Conference on Information Technology, pp.
804–809. Ieee, (2010).
[13] Juan Manuel Fern´andez, Marc Sol`a, Alexander Steblin, Eloisa Vargiu, and Felip Miralles, ‘The relevance of providing useful information to therapists and caregivers in tele*’, inDART 2014: Revised and Invited Papers; C. Lai, A. Giuliani and G. Semeraro (eds.), (in press).
[14] C´eline Franco, Jacques Demongeot, Christophe Villemazet, and Nico- las Vuillerme, ‘Behavioral telemonitoring of the elderly at home: De- tection of nycthemeral rhythms drifts from location data’, inAdvanced Information Networking and Applications Workshops (WAINA), 2010 IEEE 24th International Conference on, pp. 759–766. IEEE, (2010).
[15] Hulya Gokalp and Malcolm Clarke, ‘Monitoring activities of daily liv- ing of the elderly and the potential for its use in telecare and telehealth:
a review’,TELEMEDICINE and e-HEALTH,19(12), 910–923, (2013).
[16] Jayavardhana Gubbi, Rajkumar Buyya, Slaven Marusic, and Marimuthu Palaniswami, ‘Internet of things (iot): A vision, ar- chitectural elements, and future directions’, Future Generation Computer Systems,29(7), 1645–1660, (2013).
[17] Wu He, Gongjun Yan, and Li Da Xu, ‘Developing vehicular data cloud services in the iot environment’,Industrial Informatics, IEEE Transac- tions on,10(2), 1587–1595, (2014).
[18] Chih-Lin Hu, Hung-Tsung Huang, Cheng-Lung Lin, Nguyen Huu Minh Anh, Yi-Yu Su, and Pin-Chuan Liu, ‘Design and implemen- tation of media content sharing services in home-based iot networks’, inParallel and Distributed Systems (ICPADS), 2013 International Con- ference on, pp. 605–610. IEEE, (2013).
[19] Stephan Karpischek, Florian Michahelles, Florian Resatsch, and Elgar Fleisch, ‘Mobile sales assistant-an nfc-based product information sys- tem for retailers’, inNear Field Communication, 2009. NFC’09. First International Workshop on, pp. 20–23. IEEE, (2009).
[20] Markos Markou and Sameer Singh, ‘Novelty detection: a review?part 1:
statistical approaches’,Signal processing,83(12), 2481–2497, (2003).
[21] Felip Miralles, Eloisa Vargiu, Stefan Dauwalder, Marc Sol`a, Ger- not M¨uller-Putz, Selina C. Wriessnegger, Andreas Pinegger, Andrea K¨ubler, Sebastian Halder, Ivo K¨athner, Suzanne Martin, Jean Daly, Elaine Armstrong, Christoph Guger, Christoph Hinterm¨uller, and Han- nah Lowish, ‘Brain computer interface on track to home’,The Scientific World Journal - Advances in Brain-Computer Interface, (submitted).
[22] Felip Miralles, Eloisa Vargiu, Xavier Rafael-Palou, Marc Sol`a, Stefan Dauwalder, Christoph Guger, Christoph Hinterm¨uller, Arnau Espinosa, Hannah Lowish, Suzanne Martin, et al., ‘Brain computer interfaces on track to home: Results of the evaluation at disabled end-users’s homes and lessons learnt’,Frontiers in ICT,2, 25, (2015).
[23] Xavier Rafael-Palou, Eloisa Vargiu, Stefan Dauwalder, and Felip Mi- ralles, ‘Monitoring and supporting people that need assistance: the backhome experience’, inDART 2014: Revised and Invited Papers; C.
Lai, A. Giuliani and G. Semeraro (eds.), (in press).
[24] Xavier Rafael-Palou, Eloisa Vargiu, Guillem Serra, and Felip Miralles,
‘Improving activity monitoring through a hierarchical approach’, inThe International Conference on Information and Communication Tech- nologies for Ageing Well and e-Health (ICT 4 Ageing Well), (2015).
[25] Anand Ranganathan, Jalal Al-Muhtadi, and Roy H Campbell, ‘Rea- soning about uncertain contexts in pervasive computing environments’, IEEE Pervasive Computing,3(2), 62–70, (2004).
[26] Emmanuel Munguia Tapia, Stephen S Intille, and Kent Larson,Activity recognition in the home using simple and ubiquitous sensors, Springer, 2004.
[27] Chun-Wei Tsai, Chin-Feng Lai, and Athanasios V Vasilakos, ‘Future internet of things: open issues and challenges’, Wireless Networks, 20(8), 2201–2217, (2014).
[28] Eloisa Vargiu, Stefan Dauwalder, Xavier Rafael-Palou, Felip Miralles amd Alejandra Millet Pi-Figueres, Llu¨ısa Pla i Masip, and C`elia Ri- era Brutau, ‘Monitoring elderly people at home: Results and lessons learned’, in16th International Conference for Integrated Care 2016, Barcelona, May 23-25, (2016).
[29] Eloisa Vargiu, Juan Manuel Fern´andez, and Felip Miralles, ‘Context- aware based quality of life telemonitoring’, inDistributed Systems and Applications of Information Filtering and Retrieval. DART 2012: Re- vised and Invited Papers. C. Lai, A. Giuliani and G. Semeraro (eds.), (2014).
[30] Eloisa Vargiu, Xavier Rafael-Palou, and Felip Miralles, ‘Experimenting quality of life telemonitoring in a real scenario’,Artificial Intelligence Research,4(2), p136, (2015).
[31] Duan Yan-e, ‘Design of intelligent agriculture management information system based on iot’, inIntelligent Computation Technology and Au- tomation (ICICTA), 2011 International Conference on, volume 1, pp.
1045–1049. IEEE, (2011).
[32] Carme Zambrana, Xavier Rafael-Palou, and Eloisa Vargiu, ‘Sleeping recognition to assist elderly people at home’,Artificial Intelligence Re- search,5(2), 64–70, (2016).
Temporal goal reasoning for predictive performance of a tourist application
Eliseo Marzal, Jesus Iba˜ nez, Laura Sebastia, Eva Onaindia
1Abstract. Many real-work smart environments make use of IoT to be provided with context-aware services. Addition- ally, these environments require making decisions based on predictions about future actions. This involves the use of goal-directed behaviour which may need reasoning about new goals. This paper is devoted to analyze when a new goal can be formulated. Once a plan has been computed for a given problem, exogenous events can change the environment so that a failure in the plan is caused or an opportunity arises.
This paper present a goal reasoning framework where context information acquired from several external sources determines a change that may affect the execution of the current plan.
This change may cause a failure or an opportunity. We show how the planning system, namelyTempLM, is able to predict both failures and opportunities thanks to the analysis of the Temporal Landmarks Graph and the Temporal Propositions Graph built for the given problem.
1 INTRODUCTION
One key feature in the application of innovative technologies like IoT in smart environments is the capability of providing context-aware services. Besides real-world information, this requires anticipatory behaviour through reasoning; that is, making decisions based on predictions and expectations about future actions [1]. Particularly, many real-world applications involve unanticipated changes that may bring an alteration of the current process or a future impact on the application or even an opportunity to include some new functionality.
The purpose of a planning application is to achieve a goal through the execution of a plan or course of actions. The ar- rival of an unexpected environmental event introduces a new piece of information that was not taken into account dur- ing the construction of the plan and that may affect the ac- tive plan in several ways. Typically, the first reaction is to check if the plan is no longer executable and, if so, apply a repair mechanism or replanning to fix the failure that pre- vents the active plan from being normally executed. A second consequence is that the unanticipated event provokes a future anomaly in the plan execution. A third and more interesting derivation is whether the new data brings an opportunity to achieve a goal that is not currently contemplated in the active plan.
Goal-directed behaviour is a hallmark of intelligence aimed at discovering the changes that can be applied in the goal of
1 Universitat Polit`ecnica de Val`encia, Valencia, Spain, Email:
{emarzal,jeibrui,lstarin,onaindia}@dsic.upv.es
an application in view of the information collected by some unanticipated events. A dynamic formulation of new goals is very helpful in situations where: a) the agent’s interests are threaten and a rational anomaly response must be provided;
b) goals are no longer achievable and a graceful degradation in the goal achievement is a convenient action; or c) goal achieve- ment in the future is jeopardized, what affects future perfor- mance [14]. Thereby, goal-directed reasoning can be regarded as a context-aware responsiveness technique.
This paper is particularly devoted to analyze the first step of a goal formulation process; that is, to answer the question
’when a new goal can be formulated?’. In some applications, opportunistic behaviour is applied when the sensory input triggers some enabling conditions to accomplish a task, and reactive plans are adopted to detect problems and recover from them automatically as well as to recognize and exploit opportunities [2]. In dynamic and complex environments, like robotics, opportunities are predicted and executed immedi- ately in order to provide quick responsiveness but there is no usually anticipation of future situations.
In less dynamic and reactive environments, typically, goal formulation is considered when an anomaly is detected and/or the system is self-motivated to explore its actions in the world [14]. One approach that has been used to predict or anticipate future plan failures is Case-Based Planning (CBP). In CBP, when a plan fails, it is usually stored with the justification of its failure. This information is then retrieved from the case memory when looking for a similar situation which produced a failure in the past. CBP may be applied before the generation of a plan to anticipate possible problems and avoid situations that failed in the past, or after a plan has been produced to eliminate plans which may fail [13]. CBP presents though several limitations in its application to smart environments: a) predicting future failures is subject to finding an identical case in the case memory; and b) CBP allows only for anticipating a failure but not for detecting opportunities of pursuing a better goal or a new goal.
In this paper, we present a goal reasoning framework that traces the execution of a temporal plan and identifies if the context information acquired from several sources determines a change in the plan goals. Particularly, the reasoner detects situations of future failures and opportunities in the plan ex- ecution in the context of temporal planning with deadlines.
The framework draws uponTempLM, an approach based on temporal landmarks to handle temporal planning problems with deadlines [11, 12], and we will show how the reasoner works on a temporal plan of a tourist application.
Figure 1. Architecture
This paper is organized as follows. First, a motivating ex- ample is introduced, as well as the architecture of our system.
Then, some basic definitions referring to automated planning and the main characteristics of our plannerTempLMare given.
Section 4 introduces new definitions about exogenous events and section 5 explains the analysis thatTempLMperforms in order to detect future failures or opportunities caused by ex- ogenous events. Finally, section 6 concludes and outlines some future work.
2 TOURIST APPLICATION EXAMPLE
In order to illustrate the foundations and contributions of this paper, a problem in the context of smart tourist will be used. Smart tourisminvolves several components that are supported by information and communication technolo- gies (ICT) [9]. On one hand, it refers toSmart Destinations, which are cases of smart cities that apply smart city princi- ples to support mobility, resource availability and allocation, sustainability and quality of life/visits. Second, the Smart resident/visitor experience focuses on technology-mediated tourism experiences and their enhancement through person- alization, context-awareness and real-time monitoring [4]. Fi- nally,Smart businessrefers to the complex business ecosystem that creates and supports the exchange of touristic resources and the co-creation of the tourism experience. These smart systems include a wide range of technologies in direct sup- port of tourism such as decision support systems and the more recent recommender systems, context-aware systems, autonomous agents searching and mining Web sources, am- bient intelligence, as well as systems that create augmented realities.
In this sense, our system aims to improve the resi- dent/visitor experience by reacting in advance to changes in the environment that may cause failures or opportunities in the previously computed agenda. The architecture of our sys-
tem, shown in Figure 1, is composed of the following modules:
• TheCentral module is the core of the system. It is in charge of generating the initial plan, considering the user profile and the context information. Additionally, it listens to new events that may require to update this plan.
• The Recommender System selects the recommended visits for a tourist, given her user profile and the context information.
• The TempLM planner develops two main tasks: in first place, it receives the goals computed by the recommender system and the context information and it builds the initial plan for the tourist; then, every time a new event is received by the central module, TempLM analyses it to determine whether the plan needs to be updated.
In this paper, we will focus on the second task ofTempLM, that is, on the analysis of events to detect failures or oppor- tunities in the plan.
An example of the problem that we are facing is the follow- ing. A tourist arrives to Valencia (Spain) and she is staying at Caro Hotel. She uses our system to obtain a personalized agenda for her visit. First, the recommender system analyses her user profile to select a set of recommended visits with a recommended duration. Let us assume that the user is recom- mended to visit the Lonja for 1.5 hours, the Cathedral and the Central Market for 2 hours, respectively. These recommended goals, some other user preferences related to the time windows when she prefers to perform the activities along with infor- mation about the context, such as the opening hours of the places to visit and the geographical distances between places, are compiled into a planning problem that is formulated as an Automated Planning Problem [8], with durative actions and deadlines. This problem is solved by our plannerTempLM.
Figure 2 shows the plan obtained for this tourist. The seg- ments at the bottom represent the opening hours of places (i.e. the Lonja is open from 10h until 19h) or the time win- dows of preferences indicated by the user (i.e. the time for having lunch ranges from 14h to 16h). The green boxes rep- resent the actions in the plan. Specifically, in this domain, three types of actions can be performed: visiting an attrac- tion, having lunch and moving from one place to another.
The duration of these actions is determined by the corre- sponding parameters, that is, the attraction to visit or the ori- gin and destination of the movement action, respectively. For example, the action (visit T Lonja) takes from 10:09h to 11:29h. In addition, the visit action must consider the opening hours/preference time window; for example, the action(visit T Centralmarket)can only be performed from 15:30h until 19h. The whole plan must fit into the available time of the user indicated in Figure 2 as the timeline, that is, from 10h until 19h. A more detailed description of the compilation of this problem and domain can be found in [10].
If we consider this context, there are some events that may occur during the visit of our tourist. For example:
• The Lonja may close earlier or open later, that is, its avail- able time window may change; this may imply that the visit action has to finish before than expected or it has to be delayed, respectively.
• The Ricard Camarena restaurant may be fully-booked, which implies that another restaurant in the area must be selected.
Figure 2. Plan in execution
• The user may take longer to walk from his hotel to the Lonja, so the visit action must be delayed.
These exogenous events are received by the central mod- ule from the external data sources and are analyzed byTem- pLMin order to react in consequence. There are some events that may cause a plan failure (i.e. the attraction is closed) or a plan modification (i.e. the user gets later that expected to an attraction), whereas others may cause the appearance of a free time slot that can be used to include an affor- dance/opportunity. In this current work, we will focus on the detection of both failures and time slots for opportunities, and we will give some hints about how they can be solved or exploited.
3 BACKGROUND
This section introduces some planning concepts and then it summarizes the main characteristics of our plannerTempLM.
3.1 Planning concepts
Definition 3.1 A temporal planning problem with deadline constraints is a tupleP=
P, O, I, G, D , where P is the set of propositions,Ois the set of ground actions,I andGare sets of propositions that represent the initial state and the goal description andDis a set of deadline constraints of the form(p, t), denoting that propositionpmust be achieved withinttime units.
For example, a proposition in I can be (be T Caro), in- dicating that initially the tourist is at the Caro hotel. It is important to note that, apart from the propositions and func- tions that are initially true, the initial stateImay also contain timed initial literals (TILs). TILs, which were first introduced in PDDL2.2[6], are a very simple way of expressing that a proposition becomes true or false at a certain time point. A TIL can be represented as a pair (p, t), wherepis a (positive or negative) proposition and t is a time point. Specifically, ifpis a positive proposition, thentindicates the time point at whichp becomes true and if it is a negative proposition, thentindicates the time point at whichpbecomes false. For
example, ((not(open Lonja)),19h) is a TIL inI that indicates that the Lonja will close at 19h.
Definition 3.2 Asimple durative actiona∈Ois defined as a tuple(pre`, ef f`, pre↔, prea, ef fa, dur)[5]:
• pre`(prea) are the start (end) conditions ofa: at the state in whicha starts (ends), these conditions must hold.
• ef f`(ef fa) are the start (end) effects ofa: starting (end- ing)aupdates the world state according to these effects. A given collection of effectsef fx, x∈ {`,a}consists of:
– ef fx−, propositions to be deleted from the world state;
– ef fx+, propositions to be added to the world state
• pre↔are the invariant conditions ofa: these must hold at every point in the open interval between the start and end ofa.
• duris the duration of a.
An example of an action is shown here:
Action:Eat(?t: tourist, ?r: restaurant) pre`={(free table ?r)},prea=∅
pre↔={(open ?r), (time for eat ?t), (be ?t ?r)} ef f`=∅, ef fa={(eaten ?t)}
dur=(eat time ?t ?r)
Definition 3.3 Atemporal planΠis a set of pairs(a, t), wherea∈Oandtis the start execution time ofa.
The temporal plan for the running example is shown in Figure 2.
Definition 3.4 Given a temporal planΠ, theinduced plan Π∗for Πis the set of pairs defined as[7]:
Π∗={(a`, t),(aa, t+dur(a))}, ∀(a, t)∈Π
For simplicity, we will refer to any pair in Π∗as an ”action”.
For example, Π∗ would include ((walk T Caro Lonja)`, 10:00) and ((walk T Caro Lonja)a, 10:09). In the induced plan, we only consider the start and the end time points of the actions in the original temporal plan because these are the time points interesting for building the state resulting from the execution at a certain time pointt, as shown below.
Definition 3.5 Given a stateSt defined as a set of proposi- tions that are true at timetand a pair(ax, t)∈Π∗, an action ax isapplicableinSt if prex(a)∪pre↔(a)⊆St.
With this definition, only the start and the end time points of the actions are considered. In order to mitigate the fact that the overall conditions are not checked during the execution of the action, we check them at the start and the end of the action, although this is not consistent with the definition of a durative action in PDDL2.1 (where theoverallconditions of an actionastarting atthave to be fulfilled during the interval [t+ε, t+dur(a)−ε]).
Definition 3.6 Given a time point t, let Π∗t be the subset of actions (a, t0)∈ Π∗ such thatt0 < t. A temporal state St is composed by a set of propositions p ∈ P and a set of TILs (denoted by Γt) resulting from applying the actions in Π∗t fromI (denoted byI→Π∗t St), that is, it is assumed that the actions in Π∗t are applicable in the corresponding state.
Formally, we defineSt recursively, whereS0=I:
St=St−1−
[
∀(ax,t0)∈Π∗t
ef fx−(a)
∪
[
∀(ax,t0)∈Πt
ef fx+(a)
For example, the stateS10:09reached after the execution of the action ((walk T Caro Lonja)a,10:09h) will be the same as the initial state but S10:09 will contain the new location of the tourist (be T Lonja). It is important to notice that if we compute the stateS10:05, then the location of the user is unknown because the proposition(be T Caro)is deleted at the start time of the execution of the action and the proposition (be T Lonja)is not added until the end time of the action.
Definition 3.7 The duration (makespan) of an induced plan Π∗ executed from the initial state of the problem is dur(Π∗) =max∀(aa,t)∈Π∗ t
−Ti; i.e., the end time of the ac- tion that finishes last minus the time point at which the plan starts Ti, assuming that all the actions in Π∗ are applicable in the corresponding state.
For instance,dur(Π∗) in Figure 2 is 7 h. and 9 min., because the last action ends at 17:09h and the plan starts at 10h.
Definition 3.8 An induced plan Π∗ is a solution for a temporal planning problem with deadline constraints P=
P, O, I, G, D
if the following conditions hold:
1. G⊆Sg, whereI→Π∗t Sg, wheret=dur(Π∗) 2. ∀(p, t)∈D:∃t0≤t:p∈St0, whereI→Π∗t0 St0
This definition indicates that it is not only necessary that a plan Π∗reaches the goals, but also that all the propositions present inD are achieved before the corresponding deadline.
3.2 TempLM
TempLM [11] is a temporal planning approach for solving planning problems with deadlines that has demonstrated an excellent behaviour in the detection of unsolvable problems and the resolution of overconstrained problems. It draws upon the concept of temporal landmark, which is a proposition that
is found to necessarily happen in a solution plan in order to satisfy the deadlines of the problem goals. The set of tem- poral landmarks extracted from a problem along with their relationships form a temporal landmarks graph (TLG) that is conveniently used to take decisions during the construction of the solution plan and for guiding the search process.
Definition 3.9 ATemporal Landmarks Graph (TLG) is a directed graph G = (V, E) where the set of nodes V are landmarks and an edge in E is a tuple of the form (li, lj,≺{n,d})which represents a necessary or dependency or- dering constraint between the landmarks li and lj, denoting thatli must happen beforelj in a solution plan.
Landmarks are also annotated with varioustemporal inter- valsthat represent the validity of the corresponding temporal proposition ([11]):
• The generation interval of a landmark is denoted by [ming(l), maxg(l)]. ming(l) represents the earliest time point when landmarklcan start in the plan.maxg(l) rep- resents the latest time point whenlmust start in order to satisfyD.
• The validity interval of a landmark l is denoted by ([minv(l), maxv(l)]) and it represents the longest time that lcan be in the plan.
• The necessity interval of a landmark l is denoted by ([minn(l), maxn(l)]) and it represents the set of time points whenl is required as a condition for an action to achieve other landmarks.
These intervals are given some initial values that are then updated by means of a propagation method, as explained in [11]. In order to be consistent, both the generation interval and the necessity interval must fall into the validity interval.
Figure 3 shows a part of the initial TLG built for this prob- lem. The validity interval of a landmark is represented by a segment, themaxg is indicated by a small bold vertical line (ming is always equal tominv) and the necessity interval is represented by a green box inside the segment. For example, the validity interval of the landmark(visited T Cathedral) is [12:05h, 19h] and themaxg = 19h; the necessity interval for this landmark is empty.
TempLM applies a search process in the space of partial plans in order to find a solution plan for a problemP. A node in the search tree is defined asn = (Π, St, T LGΠ), where Π is the conflict-free partial plan ofn,Stis the state reached at timet=dur(Π) after the execution of Π fromIandT LGΠis the refined TLG after taking into account the information in Π. Given a nodenof the search tree, a successor ofnresults from adding an executable actionato the partial plan ofn, provided thata does not cause conflicts with the actions of n. Hence, the plan of the successor node will contain a newly added action, information which can be exploited to find new temporal landmarks in the TLG [12].
4 CHANGES IN THE ENVIRONMENT
This section introduces some definitions related to the man- agement of changes in the environment. We assume that in our system changes happen due to exogenous events.
